Analisis Sentimen Masyarakat terhadap Kereta Cepat Whoosh pada Platform X menggunakan IndoBERT

Abstract

The development of railroad infrastructure in Indonesia, especially high-speed rail, is relatively behind other countries. Indonesia only started the JakartaBandung high-speed rail project named Whoosh in 2015 and completed in 2023. As an Indonesia-China cooperation project with a huge investment value, this project has attracted a lot of public attention. Utilizing social media such as X as the main platform for expressing public opinion, this study aims to classify sentiment as positive, negative or neutral, in order to understand the public’s views and provide feedback for future project improvements. The complexity and volume of data from social media demands the use of artificial intelligence technologies, especially natural language processing. IndoBERT as a natural language processing model specially trained for the Indonesian language, is used in this study to effectively analyze sentiment. This research tested several pre-trained IndoBERT models from the Hugging Face website to determine the best IndoBERT model in analyzing sentiment related to the Whoosh dataset scraped on the X platform. This research also tests hyperparameter configurations such as the number of epochs, learning rate, and batch size to optimize model performance. The best model performance was achieved with a configuration of 3 epochs, learning rate 2e-5, and batch size 32, which resulted in evaluation metrics such as accuracy, recall, precision, and f1-score of 0.78. These evaluation results show that a smaller learning rate provides stable learning, while a larger batch size provides more consistent gradient estimation. However, this test shows that the model struggles to classify neutral sentiments and often misclassifies them as positive or negative. In addition, these results also show overfitting, where the model performs well on training data but degrades on testing data. This suggests that the model focuses too much on the details and noise of the training data, making it less able to generalize patterns to new data. This research highlights the importance of model selection and hyperparameter configuration in performing sentiment analysis.